Why Hosting Location Still Matters for AI Inference
"Our data is in an EU region." This is the most common response when European companies are asked about their AI provider's compliance. It is also insufficient. Data residency, jurisdiction, and sovereignty are three separate concepts, and GDPR compliance requires understanding all of them.
When your AI prompts contain customer names, financial records, or health information, the gap between "EU-hosted" and "EU-sovereign" is the gap between compliance and liability.
Three levels of location
Data residency is where the bytes physically sit. When you select AWS eu-central-1, your data resides in a Frankfurt data center. This is the easiest property to verify and the least meaningful for compliance.
Jurisdiction is which country's laws govern access to that data. AWS is a US company incorporated in Delaware. Under the CLOUD Act (2018), US authorities can compel AWS to produce data stored anywhere in the world, including Frankfurt. EU data residency does not override US jurisdiction.
Sovereignty is about operational control. Who holds the encryption keys? Can a foreign government compel the provider to access your data? Can the provider change data handling terms unilaterally? Do you have audit rights?
Most "EU hosting" discussions address residency only. For GDPR compliance with AI workloads, you need all three.
What the GDPR actually says
The relevant provisions are specific:
Article 28 (Processor obligations): Your processor must act only on documented instructions. All subprocessors must be explicitly authorized. Data must be deleted or returned when processing ends. This means you need to know exactly who touches your data and have contractual control over them.
Article 32 (Security of processing): Organizations must implement "appropriate technical and organizational measures" matching the risk level. For AI inference with personal data, this includes controlling who can access prompts and under what legal authority.
Articles 44-49 (International transfers): Any transfer of personal data outside the EEA requires specific legal mechanisms. Article 45 allows transfers to countries with adequacy decisions. The US currently has an adequacy decision under the EU-US Data Privacy Framework, but this is contested (more on this below). Article 46 permits transfers under Standard Contractual Clauses with supplementary technical measures.
Article 3 (Territorial scope): GDPR applies to any organization processing data of EU residents, regardless of where the organization is based. This means even US-based AI providers fall under GDPR when processing EU data.
The Schrems saga and why it matters for AI
In July 2020, the Court of Justice of the EU issued the Schrems II ruling, invalidating the EU-US Privacy Shield framework. The core finding: US surveillance law (FISA Section 702 and the CLOUD Act) allows mass, undisclosed access to data held by US companies, making any adequacy-based transfer mechanism unreliable.
The ruling did not ban Standard Contractual Clauses, but it required organizations to perform a "transfer impact assessment" for every transfer to a third country, documenting the legal landscape and implementing supplementary measures. The EDPB (European Data Protection Board) in Recommendations 01/2020 made this explicit: "if you can't do a credible transfer impact assessment, don't transfer the data."
In 2023, the EU adopted a new EU-US Data Privacy Framework based on US Executive Order 14086, which established a Data Protection Review Court. The EU General Court upheld this framework in September 2025. But the underlying legal risk has not changed. FISA Section 702 was reauthorized in April 2024 with expanded scope. An executive order can be revoked by any future administration. Legal scholars and privacy advocates have already called this framework "Schrems III waiting to happen."
For AI inference, this creates a specific problem. Unlike data at rest, prompts must be processed in plaintext. You cannot encrypt data in transit to the model. The only identified "adequate technical measure" per the EDPB is customer-controlled encryption with EU-held keys, but that is impossible when the provider needs to read your prompts to generate a response.
This means that for AI inference, the only reliable way to avoid transfer risk is to use an EU-incorporated provider that processes data in the EU. No transfer, no transfer impact assessment, no Schrems risk.
The cost of getting it wrong
GDPR enforcement is not theoretical. The total fines issued under GDPR exceeded EUR 4.5 billion by the end of 2025. The largest individual fines have been in the hundreds of millions. More importantly for AI workloads, multiple EU data protection authorities have issued specific guidance on AI and automated decision-making.
The practical risk is not just fines. It is the operational disruption when a DPA audit reveals that your AI inference chain sends customer data to a US company without adequate safeguards. The remediation timeline for migrating AI workloads is weeks to months, and during that time your AI-powered features are in regulatory limbo.
The EU-native provider advantage
EU-native providers (companies incorporated in the EU, operating EU data centers) eliminate the transfer question entirely. There is no international transfer. No Schrems II assessment needed. No reliance on an adequacy decision that could be invalidated.
Specific advantages:
| Property | US provider (EU region) | EU-native provider |
|---|---|---|
| Data residency | EU (data center in EU) | EU |
| Jurisdiction | US (CLOUD Act applies) | EU (local law only) |
| Sovereignty | Limited (US parent controls access) | Full (EU entity, EU legal framework) |
| SEAL level | SEAL-1 | SEAL-2 or SEAL-3 |
| Transfer impact assessment | Required | Not applicable |
| Schrems III risk | Exposed | None |
There is also a cost advantage. According to analysis from February 2026, EU-native cloud providers like Scaleway and OVHcloud offer 4 to 14 times the compute value per euro compared to US hyperscalers. The compliance argument and the economic argument point in the same direction.
What this means for AI inference specifically
AI inference is different from most cloud workloads in three important ways:
1. Prompts contain the most sensitive data. When a financial services company sends a prompt like "Summarize the credit history of [Customer Name] and recommend an approval decision," the prompt itself contains personal data. This is not metadata or aggregate statistics. It is the raw data.
2. Data must be processed in plaintext. Unlike storage or transport, inference requires the provider to read your data. Encryption at rest and in transit does not protect data during processing (unless you use confidential computing, which is SEAL-4 and not widely available for AI workloads yet).
3. Retention creates compounding risk. If your provider stores prompts for model improvement, abuse detection, or debugging, every prompt ever sent becomes data at risk in a breach. Zero Data Retention is not just a privacy feature. It is a risk reduction mechanism.
ozeye's approach
ozeye uses three EU-native providers: Mistral AI (Paris, France), Scaleway (Paris, France), and OVHcloud (Gravelines, France). All three are French-incorporated companies operating French data centers. No US parent companies. No CLOUD Act exposure.
All three operate with Zero Data Retention for API inference. Prompts and completions are not stored after the response is returned. The only data retained is billing metadata: model, token count, cost, timestamp. This means:
- No international transfer under GDPR Articles 44-49
- No transfer impact assessment required
- No exposure to Schrems III risk
- No stored prompts to compromise in a breach
- Complete subprocessor transparency
If you are currently sending EU personal data through a US AI provider, switching to ozeye is a configuration change. Same OpenAI-compatible API, same SDKs, different base URL. The compliance improvement is immediate.
Three questions to ask your current provider
Before your next compliance review, ask your AI provider:
1. Where is the company incorporated? Not where the data center is. Where the legal entity is. If the answer is the United States, your data is subject to US jurisdiction regardless of hosting location.
2. Who are all the subprocessors? Every service that touches your data, including monitoring, logging, and abuse detection. EDPB Opinion 22/2024 requires real-time visibility into the full processor chain.
3. What data is retained, and for how long? If prompts are stored for any purpose, even temporarily, that is data at risk. The minimum retention period should be zero for inference workloads.
If you do not like the answers, the alternative is straightforward.